Quantum computing advances universal function derivation with neural networks and embedding techniques

The challenge of accurately simulating the behavior of interacting particles is at the heart of many problems in physics and materials science, requiring increasingly powerful computational techniques. Martin J. Uttendorfer, Daniel Barragan-Yani, Matthias Sperl, and Marc Landmann, all from the German Aerospace Center, present an innovative approach that integrates quantum computing, neural networks, and density matrix embedding theory to tackle this complex problem. Their research focuses on deriving “universal functionals,” mathematical tools that simplify these simulations by training deep neural networks using advanced algorithms. By incorporating the fragment bus system, the team significantly expands the range of physical systems to which this functionality can be applied, potentially unlocking the cumulative benefits of computational power in a variety of applications in condensate physics and beyond.

Their research focuses on deriving “universal functionals,” mathematical tools that simplify these simulations by training deep neural networks using advanced algorithms. By incorporating the fragment bus system, the team significantly expands the range of physical systems to which this function can be applied, potentially unlocking cumulative benefits of computational power in a variety of applications in condensate physics and beyond.

Neural networks unlock universal particle interactions

Researchers have developed a new method that integrates computing, machine learning, and reduced density matrix functional theory to improve simulations of interacting particles. The core of this research lies in using deep neural networks trained with advanced algorithms to obtain universal functionals, which are mathematical descriptions of particle interactions. To expand the applicability of this function, researchers adopted a fragment bus system, which significantly expanded the range of systems to which it can be applied. The researchers' approach uses Lagrangian multipliers to define the Hamiltonian by scanning various one-electron-reduced density matrices, circumventing traditional constraints and allowing function values ​​to be calculated through mathematical transformations.

This process involves iteratively finding the ground state energies of different Hamiltonians and converting them into functional forms to effectively train machine learning models. Measurements confirm that this method overcomes the limitations of traditional approaches, as the resulting deep neural network exhibits a linear gradient even for complex Hamiltonian terms, simplifying the learning process. Experiments show that deep neural network functionals trained on data generated from quantum computation can accurately represent particle interactions even in regions where standard density functional theory is difficult. The team utilized a fully connected multilayer neural network and demonstrated that with sufficient training data, a complex network can achieve energy accuracy comparable to high-precision computation.

Importantly, the obtained results show that minor errors in quantum processors only lead to minor inaccuracies in machine learning models. This method naturally leverages every computational run to train the network, providing significant advantages over methods that require iterative tuning and allowing for significant parallelism. Furthermore, this approach inherently guarantees v-representability. This means that the resulting function accurately describes physically possible states.

Machine learning speeds up density matrix simulations

This study demonstrates a new approach to simulating interacting particles by integrating reduced density matrix functional theory and machine learning techniques. Scientists have successfully trained deep neural networks to obtain universal functions, a key step toward more efficient and accurate simulations in condensed matter physics and quantum chemistry. This method exploits the fragment bus system, expands the range of applicable Hamiltonians, and provides potential cumulative benefits in computational workflows. The team's work yields more computationally efficient and accurate results compared to existing quantum chemistry methods of similar cost, especially for lattice models where the scaling is cubic with system size.

Incorporating density matrix embedding theory enhanced the versatility of the function and justified the initial computational cost required to train the data. This approach also provides scalability as training data can be generated through parallel processes, reducing dependence on quantum processing units. This is a valuable advantage considering the limited availability of quantum hardware. The researchers acknowledge that extending this scheme to molecular systems can result in higher computational costs associated with orbital localization and orthogonalization. However, they show that future work can easily extend this scheme to orbital-based quantum chemical calculations, building on the current framework and potentially providing improvements over coupled cluster methods in terms of computational scaling.